Stats
Links
Categories
Memibrium — Sovereign AI Memory with Crystallization Theory
"It gets smarter without getting wiser." — That's the problem this stack is built to solve.
Sovereign, self-hosted memory infrastructure implementing Crystallization Theory as a knowledge governance layer over tiered vector search. Ships as a plugin: one skill + one MCP server = governed memory for any agent.
Patent POC: CT #63/953,509 · KEOS #63/962,609 · STG (pending)
Install in 60 Seconds
Prerequisites
- Python 3.11+, PostgreSQL with pgvector extension
- Any OpenAI-compatible LLM provider (OpenAI, Azure, Ollama, etc.)
- Node.js 18+ (for plugin hosts that use npx)
1. Start the server
pip install asyncpg openai starlette uvicorn
export OPENAI_API_KEY="your-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"
export DB_HOST=localhost DB_NAME=memory DB_USER=memory DB_PASSWORD=memory
python server.py
2. Install the plugin
Claude Code:
claude plugin marketplace add rvalen1123/memibrium
claude plugin install memibrium@memibrium
Copilot CLI:
/plugin marketplace add rvalen1123/memibrium
/plugin install memibrium@memibrium
Cursor / other MCP clients: See DISTRIBUTION.md
3. Verify
Ask your agent: "What do you know about my project?"
You should get a governed memory response with lifecycle states — not a generic "I don't have memory" answer.
One Install, Two Layers
The skill: the brain
The plugin ships one governance skill (crystallization-memory) that teaches agents how memory works in this system. It provides 8 behavioral rules, anti-patterns, and interaction patterns:
- All knowledge enters as untrusted (OBSERVATION state)
- Check lifecycle state before citing any memory
- Human confirmation is the ONLY crystallization path
- Contradictions trigger investigation, not overwrites
- Freeze before destructive changes; revert if wrong
The MCP server: the hands
The plugin wires in the Memibrium MCP Server, which gives your agent 8 tools for governed memory operations. That is the execution layer for storing, retrieving, confirming, freezing, and reverting knowledge.
Why this plugin is different
This is not a prompt pack. It is a packaged knowledge governance layer:
- The skill teaches the agent when to store, when to cite, and what to avoid.
- The MCP tools let the agent act on live memory with lifecycle guarantees.
- The plugin keeps the guidance layer and execution layer aligned in one install.
- Multi-host support lets you use the same governed memory across Claude Code, Copilot CLI, Cursor, and other compatible hosts.
Architecture
Plugin Install
├─ Skill (SKILL.md) → loaded into agent system prompt
│ "How the agent should think about memory"
│
└─ MCP Server (.mcp.json → server.py on :9999)
└─ CT Lifecycle Engine (the patent layer)
└─ pgvector dual-tier (hot=working, cold=crystallized)
└─ Any OpenAI-compatible LLM provider
Tiering policy = lifecycle state. That's the paper.
| Tier | States | Engine | Latency |
|---|---|---|---|
| Hot | observation, consideration, accepted | RuVector HNSW + GNN (or pgvector fallback) | <1ms |
| Cold | crystallized, shed | LEANN graph recomputation (or pgvector fallback) | ~250ms |
What You Get
| Component | What it adds | Examples |
|---|---|---|
| Crystallization skill | Memory governance, behavioral rules, guardrails | 8 rules: cite by state, never auto-crystallize, freeze before destructive changes |
| Memibrium MCP | Live memory tooling with lifecycle guarantees | retain, recall, reflect, confirm, freeze, revert, consolidate, dashboard |
| RuVector engine | GNN re-ranking + SONA self-learning on hot tier | Sub-millisecond HNSW, results improve over time, drop-in pgvector replacement |
| LEANN cold tier | 97% storage compression on crystallized memories | Graph-based recomputation, no stored embeddings, ~250ms search |